Hallucination
A hallucination is output from an AI system that is fluent and confident but not grounded in the input, invented, added or contradicted. In translation it appears as content the source never contained, a mistranslation stated with full assurance, or detail the model supplied on its own. It is the central reason machine translation output cannot be trusted unchecked.
How it works
Generative models produce the most plausible-sounding continuation, not the most accurate one. When the source is ambiguous, unusual or outside what the model saw in training, the most plausible output can diverge from the truth, and the model has no internal signal that it has done so.
The danger is that hallucinations are fluent by construction. Nothing on the surface distinguishes an invented sentence from a correct one, so they pass casual reading and only a checker with the source catches them.
How SourceTarget uses it
SourceTarget's confidence and quality scoring exist partly to surface likely hallucinations: segments where the engine is least certain, or where output diverges from expectation, are flagged for the post-editor. Human review with the source in hand is the actual safeguard.